xref: /freebsd/sys/kern/subr_taskqueue.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2000 Doug Rabson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/bus.h>
35 #include <sys/cpuset.h>
36 #include <sys/interrupt.h>
37 #include <sys/kernel.h>
38 #include <sys/kthread.h>
39 #include <sys/libkern.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/mutex.h>
44 #include <sys/proc.h>
45 #include <sys/sched.h>
46 #include <sys/smp.h>
47 #include <sys/taskqueue.h>
48 #include <sys/unistd.h>
49 #include <machine/stdarg.h>
50 
51 static MALLOC_DEFINE(M_TASKQUEUE, "taskqueue", "Task Queues");
52 static void	*taskqueue_giant_ih;
53 static void	*taskqueue_ih;
54 static void	 taskqueue_fast_enqueue(void *);
55 static void	 taskqueue_swi_enqueue(void *);
56 static void	 taskqueue_swi_giant_enqueue(void *);
57 
58 struct taskqueue_busy {
59 	struct task	*tb_running;
60 	TAILQ_ENTRY(taskqueue_busy) tb_link;
61 };
62 
63 struct task * const TB_DRAIN_WAITER = (struct task *)0x1;
64 
65 struct taskqueue {
66 	STAILQ_HEAD(, task)	tq_queue;
67 	taskqueue_enqueue_fn	tq_enqueue;
68 	void			*tq_context;
69 	char			*tq_name;
70 	TAILQ_HEAD(, taskqueue_busy) tq_active;
71 	struct mtx		tq_mutex;
72 	struct thread		**tq_threads;
73 	int			tq_tcount;
74 	int			tq_spin;
75 	int			tq_flags;
76 	int			tq_callouts;
77 	taskqueue_callback_fn	tq_callbacks[TASKQUEUE_NUM_CALLBACKS];
78 	void			*tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS];
79 };
80 
81 #define	TQ_FLAGS_ACTIVE		(1 << 0)
82 #define	TQ_FLAGS_BLOCKED	(1 << 1)
83 #define	TQ_FLAGS_UNLOCKED_ENQUEUE	(1 << 2)
84 
85 #define	DT_CALLOUT_ARMED	(1 << 0)
86 #define	DT_DRAIN_IN_PROGRESS	(1 << 1)
87 
88 #define	TQ_LOCK(tq)							\
89 	do {								\
90 		if ((tq)->tq_spin)					\
91 			mtx_lock_spin(&(tq)->tq_mutex);			\
92 		else							\
93 			mtx_lock(&(tq)->tq_mutex);			\
94 	} while (0)
95 #define	TQ_ASSERT_LOCKED(tq)	mtx_assert(&(tq)->tq_mutex, MA_OWNED)
96 
97 #define	TQ_UNLOCK(tq)							\
98 	do {								\
99 		if ((tq)->tq_spin)					\
100 			mtx_unlock_spin(&(tq)->tq_mutex);		\
101 		else							\
102 			mtx_unlock(&(tq)->tq_mutex);			\
103 	} while (0)
104 #define	TQ_ASSERT_UNLOCKED(tq)	mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED)
105 
106 void
107 _timeout_task_init(struct taskqueue *queue, struct timeout_task *timeout_task,
108     int priority, task_fn_t func, void *context)
109 {
110 
111 	TASK_INIT(&timeout_task->t, priority, func, context);
112 	callout_init_mtx(&timeout_task->c, &queue->tq_mutex,
113 	    CALLOUT_RETURNUNLOCKED);
114 	timeout_task->q = queue;
115 	timeout_task->f = 0;
116 }
117 
118 static __inline int
119 TQ_SLEEP(struct taskqueue *tq, void *p, struct mtx *m, int pri, const char *wm,
120     int t)
121 {
122 	if (tq->tq_spin)
123 		return (msleep_spin(p, m, wm, t));
124 	return (msleep(p, m, pri, wm, t));
125 }
126 
127 static struct taskqueue *
128 _taskqueue_create(const char *name, int mflags,
129 		 taskqueue_enqueue_fn enqueue, void *context,
130 		 int mtxflags, const char *mtxname __unused)
131 {
132 	struct taskqueue *queue;
133 	char *tq_name;
134 
135 	tq_name = malloc(TASKQUEUE_NAMELEN, M_TASKQUEUE, mflags | M_ZERO);
136 	if (tq_name == NULL)
137 		return (NULL);
138 
139 	queue = malloc(sizeof(struct taskqueue), M_TASKQUEUE, mflags | M_ZERO);
140 	if (queue == NULL) {
141 		free(tq_name, M_TASKQUEUE);
142 		return (NULL);
143 	}
144 
145 	snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue");
146 
147 	STAILQ_INIT(&queue->tq_queue);
148 	TAILQ_INIT(&queue->tq_active);
149 	queue->tq_enqueue = enqueue;
150 	queue->tq_context = context;
151 	queue->tq_name = tq_name;
152 	queue->tq_spin = (mtxflags & MTX_SPIN) != 0;
153 	queue->tq_flags |= TQ_FLAGS_ACTIVE;
154 	if (enqueue == taskqueue_fast_enqueue ||
155 	    enqueue == taskqueue_swi_enqueue ||
156 	    enqueue == taskqueue_swi_giant_enqueue ||
157 	    enqueue == taskqueue_thread_enqueue)
158 		queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE;
159 	mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags);
160 
161 	return (queue);
162 }
163 
164 struct taskqueue *
165 taskqueue_create(const char *name, int mflags,
166 		 taskqueue_enqueue_fn enqueue, void *context)
167 {
168 
169 	return _taskqueue_create(name, mflags, enqueue, context,
170 			MTX_DEF, name);
171 }
172 
173 void
174 taskqueue_set_callback(struct taskqueue *queue,
175     enum taskqueue_callback_type cb_type, taskqueue_callback_fn callback,
176     void *context)
177 {
178 
179 	KASSERT(((cb_type >= TASKQUEUE_CALLBACK_TYPE_MIN) &&
180 	    (cb_type <= TASKQUEUE_CALLBACK_TYPE_MAX)),
181 	    ("Callback type %d not valid, must be %d-%d", cb_type,
182 	    TASKQUEUE_CALLBACK_TYPE_MIN, TASKQUEUE_CALLBACK_TYPE_MAX));
183 	KASSERT((queue->tq_callbacks[cb_type] == NULL),
184 	    ("Re-initialization of taskqueue callback?"));
185 
186 	queue->tq_callbacks[cb_type] = callback;
187 	queue->tq_cb_contexts[cb_type] = context;
188 }
189 
190 /*
191  * Signal a taskqueue thread to terminate.
192  */
193 static void
194 taskqueue_terminate(struct thread **pp, struct taskqueue *tq)
195 {
196 
197 	while (tq->tq_tcount > 0 || tq->tq_callouts > 0) {
198 		wakeup(tq);
199 		TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0);
200 	}
201 }
202 
203 void
204 taskqueue_free(struct taskqueue *queue)
205 {
206 
207 	TQ_LOCK(queue);
208 	queue->tq_flags &= ~TQ_FLAGS_ACTIVE;
209 	taskqueue_terminate(queue->tq_threads, queue);
210 	KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?"));
211 	KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks"));
212 	mtx_destroy(&queue->tq_mutex);
213 	free(queue->tq_threads, M_TASKQUEUE);
214 	free(queue->tq_name, M_TASKQUEUE);
215 	free(queue, M_TASKQUEUE);
216 }
217 
218 static int
219 taskqueue_enqueue_locked(struct taskqueue *queue, struct task *task)
220 {
221 	struct task *ins;
222 	struct task *prev;
223 
224 	KASSERT(task->ta_func != NULL, ("enqueueing task with NULL func"));
225 	/*
226 	 * Count multiple enqueues.
227 	 */
228 	if (task->ta_pending) {
229 		if (task->ta_pending < USHRT_MAX)
230 			task->ta_pending++;
231 		TQ_UNLOCK(queue);
232 		return (0);
233 	}
234 
235 	/*
236 	 * Optimise the case when all tasks have the same priority.
237 	 */
238 	prev = STAILQ_LAST(&queue->tq_queue, task, ta_link);
239 	if (!prev || prev->ta_priority >= task->ta_priority) {
240 		STAILQ_INSERT_TAIL(&queue->tq_queue, task, ta_link);
241 	} else {
242 		prev = NULL;
243 		for (ins = STAILQ_FIRST(&queue->tq_queue); ins;
244 		     prev = ins, ins = STAILQ_NEXT(ins, ta_link))
245 			if (ins->ta_priority < task->ta_priority)
246 				break;
247 
248 		if (prev)
249 			STAILQ_INSERT_AFTER(&queue->tq_queue, prev, task, ta_link);
250 		else
251 			STAILQ_INSERT_HEAD(&queue->tq_queue, task, ta_link);
252 	}
253 
254 	task->ta_pending = 1;
255 	if ((queue->tq_flags & TQ_FLAGS_UNLOCKED_ENQUEUE) != 0)
256 		TQ_UNLOCK(queue);
257 	if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0)
258 		queue->tq_enqueue(queue->tq_context);
259 	if ((queue->tq_flags & TQ_FLAGS_UNLOCKED_ENQUEUE) == 0)
260 		TQ_UNLOCK(queue);
261 
262 	/* Return with lock released. */
263 	return (0);
264 }
265 
266 int
267 taskqueue_enqueue(struct taskqueue *queue, struct task *task)
268 {
269 	int res;
270 
271 	TQ_LOCK(queue);
272 	res = taskqueue_enqueue_locked(queue, task);
273 	/* The lock is released inside. */
274 
275 	return (res);
276 }
277 
278 static void
279 taskqueue_timeout_func(void *arg)
280 {
281 	struct taskqueue *queue;
282 	struct timeout_task *timeout_task;
283 
284 	timeout_task = arg;
285 	queue = timeout_task->q;
286 	KASSERT((timeout_task->f & DT_CALLOUT_ARMED) != 0, ("Stray timeout"));
287 	timeout_task->f &= ~DT_CALLOUT_ARMED;
288 	queue->tq_callouts--;
289 	taskqueue_enqueue_locked(timeout_task->q, &timeout_task->t);
290 	/* The lock is released inside. */
291 }
292 
293 int
294 taskqueue_enqueue_timeout_sbt(struct taskqueue *queue,
295     struct timeout_task *timeout_task, sbintime_t sbt, sbintime_t pr, int flags)
296 {
297 	int res;
298 
299 	TQ_LOCK(queue);
300 	KASSERT(timeout_task->q == NULL || timeout_task->q == queue,
301 	    ("Migrated queue"));
302 	KASSERT(!queue->tq_spin, ("Timeout for spin-queue"));
303 	timeout_task->q = queue;
304 	res = timeout_task->t.ta_pending;
305 	if (timeout_task->f & DT_DRAIN_IN_PROGRESS) {
306 		/* Do nothing */
307 		TQ_UNLOCK(queue);
308 		res = -1;
309 	} else if (sbt == 0) {
310 		taskqueue_enqueue_locked(queue, &timeout_task->t);
311 		/* The lock is released inside. */
312 	} else {
313 		if ((timeout_task->f & DT_CALLOUT_ARMED) != 0) {
314 			res++;
315 		} else {
316 			queue->tq_callouts++;
317 			timeout_task->f |= DT_CALLOUT_ARMED;
318 			if (sbt < 0)
319 				sbt = -sbt; /* Ignore overflow. */
320 		}
321 		if (sbt > 0) {
322 			callout_reset_sbt(&timeout_task->c, sbt, pr,
323 			    taskqueue_timeout_func, timeout_task, flags);
324 		}
325 		TQ_UNLOCK(queue);
326 	}
327 	return (res);
328 }
329 
330 int
331 taskqueue_enqueue_timeout(struct taskqueue *queue,
332     struct timeout_task *ttask, int ticks)
333 {
334 
335 	return (taskqueue_enqueue_timeout_sbt(queue, ttask, ticks * tick_sbt,
336 	    0, 0));
337 }
338 
339 static void
340 taskqueue_task_nop_fn(void *context, int pending)
341 {
342 }
343 
344 /*
345  * Block until all currently queued tasks in this taskqueue
346  * have begun execution.  Tasks queued during execution of
347  * this function are ignored.
348  */
349 static int
350 taskqueue_drain_tq_queue(struct taskqueue *queue)
351 {
352 	struct task t_barrier;
353 
354 	if (STAILQ_EMPTY(&queue->tq_queue))
355 		return (0);
356 
357 	/*
358 	 * Enqueue our barrier after all current tasks, but with
359 	 * the highest priority so that newly queued tasks cannot
360 	 * pass it.  Because of the high priority, we can not use
361 	 * taskqueue_enqueue_locked directly (which drops the lock
362 	 * anyway) so just insert it at tail while we have the
363 	 * queue lock.
364 	 */
365 	TASK_INIT(&t_barrier, USHRT_MAX, taskqueue_task_nop_fn, &t_barrier);
366 	STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link);
367 	t_barrier.ta_pending = 1;
368 
369 	/*
370 	 * Once the barrier has executed, all previously queued tasks
371 	 * have completed or are currently executing.
372 	 */
373 	while (t_barrier.ta_pending != 0)
374 		TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0);
375 	return (1);
376 }
377 
378 /*
379  * Block until all currently executing tasks for this taskqueue
380  * complete.  Tasks that begin execution during the execution
381  * of this function are ignored.
382  */
383 static int
384 taskqueue_drain_tq_active(struct taskqueue *queue)
385 {
386 	struct taskqueue_busy tb_marker, *tb_first;
387 
388 	if (TAILQ_EMPTY(&queue->tq_active))
389 		return (0);
390 
391 	/* Block taskq_terminate().*/
392 	queue->tq_callouts++;
393 
394 	/*
395 	 * Wait for all currently executing taskqueue threads
396 	 * to go idle.
397 	 */
398 	tb_marker.tb_running = TB_DRAIN_WAITER;
399 	TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link);
400 	while (TAILQ_FIRST(&queue->tq_active) != &tb_marker)
401 		TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0);
402 	TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link);
403 
404 	/*
405 	 * Wakeup any other drain waiter that happened to queue up
406 	 * without any intervening active thread.
407 	 */
408 	tb_first = TAILQ_FIRST(&queue->tq_active);
409 	if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER)
410 		wakeup(tb_first);
411 
412 	/* Release taskqueue_terminate(). */
413 	queue->tq_callouts--;
414 	if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0)
415 		wakeup_one(queue->tq_threads);
416 	return (1);
417 }
418 
419 void
420 taskqueue_block(struct taskqueue *queue)
421 {
422 
423 	TQ_LOCK(queue);
424 	queue->tq_flags |= TQ_FLAGS_BLOCKED;
425 	TQ_UNLOCK(queue);
426 }
427 
428 void
429 taskqueue_unblock(struct taskqueue *queue)
430 {
431 
432 	TQ_LOCK(queue);
433 	queue->tq_flags &= ~TQ_FLAGS_BLOCKED;
434 	if (!STAILQ_EMPTY(&queue->tq_queue))
435 		queue->tq_enqueue(queue->tq_context);
436 	TQ_UNLOCK(queue);
437 }
438 
439 static void
440 taskqueue_run_locked(struct taskqueue *queue)
441 {
442 	struct taskqueue_busy tb;
443 	struct taskqueue_busy *tb_first;
444 	struct task *task;
445 	int pending;
446 
447 	KASSERT(queue != NULL, ("tq is NULL"));
448 	TQ_ASSERT_LOCKED(queue);
449 	tb.tb_running = NULL;
450 
451 	while (STAILQ_FIRST(&queue->tq_queue)) {
452 		TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link);
453 
454 		/*
455 		 * Carefully remove the first task from the queue and
456 		 * zero its pending count.
457 		 */
458 		task = STAILQ_FIRST(&queue->tq_queue);
459 		KASSERT(task != NULL, ("task is NULL"));
460 		STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link);
461 		pending = task->ta_pending;
462 		task->ta_pending = 0;
463 		tb.tb_running = task;
464 		TQ_UNLOCK(queue);
465 
466 		KASSERT(task->ta_func != NULL, ("task->ta_func is NULL"));
467 		task->ta_func(task->ta_context, pending);
468 
469 		TQ_LOCK(queue);
470 		tb.tb_running = NULL;
471 		wakeup(task);
472 
473 		TAILQ_REMOVE(&queue->tq_active, &tb, tb_link);
474 		tb_first = TAILQ_FIRST(&queue->tq_active);
475 		if (tb_first != NULL &&
476 		    tb_first->tb_running == TB_DRAIN_WAITER)
477 			wakeup(tb_first);
478 	}
479 }
480 
481 void
482 taskqueue_run(struct taskqueue *queue)
483 {
484 
485 	TQ_LOCK(queue);
486 	taskqueue_run_locked(queue);
487 	TQ_UNLOCK(queue);
488 }
489 
490 static int
491 task_is_running(struct taskqueue *queue, struct task *task)
492 {
493 	struct taskqueue_busy *tb;
494 
495 	TQ_ASSERT_LOCKED(queue);
496 	TAILQ_FOREACH(tb, &queue->tq_active, tb_link) {
497 		if (tb->tb_running == task)
498 			return (1);
499 	}
500 	return (0);
501 }
502 
503 /*
504  * Only use this function in single threaded contexts. It returns
505  * non-zero if the given task is either pending or running. Else the
506  * task is idle and can be queued again or freed.
507  */
508 int
509 taskqueue_poll_is_busy(struct taskqueue *queue, struct task *task)
510 {
511 	int retval;
512 
513 	TQ_LOCK(queue);
514 	retval = task->ta_pending > 0 || task_is_running(queue, task);
515 	TQ_UNLOCK(queue);
516 
517 	return (retval);
518 }
519 
520 static int
521 taskqueue_cancel_locked(struct taskqueue *queue, struct task *task,
522     u_int *pendp)
523 {
524 
525 	if (task->ta_pending > 0)
526 		STAILQ_REMOVE(&queue->tq_queue, task, task, ta_link);
527 	if (pendp != NULL)
528 		*pendp = task->ta_pending;
529 	task->ta_pending = 0;
530 	return (task_is_running(queue, task) ? EBUSY : 0);
531 }
532 
533 int
534 taskqueue_cancel(struct taskqueue *queue, struct task *task, u_int *pendp)
535 {
536 	int error;
537 
538 	TQ_LOCK(queue);
539 	error = taskqueue_cancel_locked(queue, task, pendp);
540 	TQ_UNLOCK(queue);
541 
542 	return (error);
543 }
544 
545 int
546 taskqueue_cancel_timeout(struct taskqueue *queue,
547     struct timeout_task *timeout_task, u_int *pendp)
548 {
549 	u_int pending, pending1;
550 	int error;
551 
552 	TQ_LOCK(queue);
553 	pending = !!(callout_stop(&timeout_task->c) > 0);
554 	error = taskqueue_cancel_locked(queue, &timeout_task->t, &pending1);
555 	if ((timeout_task->f & DT_CALLOUT_ARMED) != 0) {
556 		timeout_task->f &= ~DT_CALLOUT_ARMED;
557 		queue->tq_callouts--;
558 	}
559 	TQ_UNLOCK(queue);
560 
561 	if (pendp != NULL)
562 		*pendp = pending + pending1;
563 	return (error);
564 }
565 
566 void
567 taskqueue_drain(struct taskqueue *queue, struct task *task)
568 {
569 
570 	if (!queue->tq_spin)
571 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
572 
573 	TQ_LOCK(queue);
574 	while (task->ta_pending != 0 || task_is_running(queue, task))
575 		TQ_SLEEP(queue, task, &queue->tq_mutex, PWAIT, "-", 0);
576 	TQ_UNLOCK(queue);
577 }
578 
579 void
580 taskqueue_drain_all(struct taskqueue *queue)
581 {
582 
583 	if (!queue->tq_spin)
584 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__);
585 
586 	TQ_LOCK(queue);
587 	(void)taskqueue_drain_tq_queue(queue);
588 	(void)taskqueue_drain_tq_active(queue);
589 	TQ_UNLOCK(queue);
590 }
591 
592 void
593 taskqueue_drain_timeout(struct taskqueue *queue,
594     struct timeout_task *timeout_task)
595 {
596 
597 	/*
598 	 * Set flag to prevent timer from re-starting during drain:
599 	 */
600 	TQ_LOCK(queue);
601 	KASSERT((timeout_task->f & DT_DRAIN_IN_PROGRESS) == 0,
602 	    ("Drain already in progress"));
603 	timeout_task->f |= DT_DRAIN_IN_PROGRESS;
604 	TQ_UNLOCK(queue);
605 
606 	callout_drain(&timeout_task->c);
607 	taskqueue_drain(queue, &timeout_task->t);
608 
609 	/*
610 	 * Clear flag to allow timer to re-start:
611 	 */
612 	TQ_LOCK(queue);
613 	timeout_task->f &= ~DT_DRAIN_IN_PROGRESS;
614 	TQ_UNLOCK(queue);
615 }
616 
617 void
618 taskqueue_quiesce(struct taskqueue *queue)
619 {
620 	int ret;
621 
622 	TQ_LOCK(queue);
623 	do {
624 		ret = taskqueue_drain_tq_queue(queue);
625 		if (ret == 0)
626 			ret = taskqueue_drain_tq_active(queue);
627 	} while (ret != 0);
628 	TQ_UNLOCK(queue);
629 }
630 
631 static void
632 taskqueue_swi_enqueue(void *context)
633 {
634 	swi_sched(taskqueue_ih, 0);
635 }
636 
637 static void
638 taskqueue_swi_run(void *dummy)
639 {
640 	taskqueue_run(taskqueue_swi);
641 }
642 
643 static void
644 taskqueue_swi_giant_enqueue(void *context)
645 {
646 	swi_sched(taskqueue_giant_ih, 0);
647 }
648 
649 static void
650 taskqueue_swi_giant_run(void *dummy)
651 {
652 	taskqueue_run(taskqueue_swi_giant);
653 }
654 
655 static int
656 _taskqueue_start_threads(struct taskqueue **tqp, int count, int pri,
657     cpuset_t *mask, const char *name, va_list ap)
658 {
659 	char ktname[MAXCOMLEN + 1];
660 	struct thread *td;
661 	struct taskqueue *tq;
662 	int i, error;
663 
664 	if (count <= 0)
665 		return (EINVAL);
666 
667 	vsnprintf(ktname, sizeof(ktname), name, ap);
668 	tq = *tqp;
669 
670 	tq->tq_threads = malloc(sizeof(struct thread *) * count, M_TASKQUEUE,
671 	    M_NOWAIT | M_ZERO);
672 	if (tq->tq_threads == NULL) {
673 		printf("%s: no memory for %s threads\n", __func__, ktname);
674 		return (ENOMEM);
675 	}
676 
677 	for (i = 0; i < count; i++) {
678 		if (count == 1)
679 			error = kthread_add(taskqueue_thread_loop, tqp, NULL,
680 			    &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname);
681 		else
682 			error = kthread_add(taskqueue_thread_loop, tqp, NULL,
683 			    &tq->tq_threads[i], RFSTOPPED, 0,
684 			    "%s_%d", ktname, i);
685 		if (error) {
686 			/* should be ok to continue, taskqueue_free will dtrt */
687 			printf("%s: kthread_add(%s): error %d", __func__,
688 			    ktname, error);
689 			tq->tq_threads[i] = NULL;		/* paranoid */
690 		} else
691 			tq->tq_tcount++;
692 	}
693 	if (tq->tq_tcount == 0) {
694 		free(tq->tq_threads, M_TASKQUEUE);
695 		tq->tq_threads = NULL;
696 		return (ENOMEM);
697 	}
698 	for (i = 0; i < count; i++) {
699 		if (tq->tq_threads[i] == NULL)
700 			continue;
701 		td = tq->tq_threads[i];
702 		if (mask) {
703 			error = cpuset_setthread(td->td_tid, mask);
704 			/*
705 			 * Failing to pin is rarely an actual fatal error;
706 			 * it'll just affect performance.
707 			 */
708 			if (error)
709 				printf("%s: curthread=%llu: can't pin; "
710 				    "error=%d\n",
711 				    __func__,
712 				    (unsigned long long) td->td_tid,
713 				    error);
714 		}
715 		thread_lock(td);
716 		sched_prio(td, pri);
717 		sched_add(td, SRQ_BORING);
718 		thread_unlock(td);
719 	}
720 
721 	return (0);
722 }
723 
724 int
725 taskqueue_start_threads(struct taskqueue **tqp, int count, int pri,
726     const char *name, ...)
727 {
728 	va_list ap;
729 	int error;
730 
731 	va_start(ap, name);
732 	error = _taskqueue_start_threads(tqp, count, pri, NULL, name, ap);
733 	va_end(ap);
734 	return (error);
735 }
736 
737 int
738 taskqueue_start_threads_cpuset(struct taskqueue **tqp, int count, int pri,
739     cpuset_t *mask, const char *name, ...)
740 {
741 	va_list ap;
742 	int error;
743 
744 	va_start(ap, name);
745 	error = _taskqueue_start_threads(tqp, count, pri, mask, name, ap);
746 	va_end(ap);
747 	return (error);
748 }
749 
750 static inline void
751 taskqueue_run_callback(struct taskqueue *tq,
752     enum taskqueue_callback_type cb_type)
753 {
754 	taskqueue_callback_fn tq_callback;
755 
756 	TQ_ASSERT_UNLOCKED(tq);
757 	tq_callback = tq->tq_callbacks[cb_type];
758 	if (tq_callback != NULL)
759 		tq_callback(tq->tq_cb_contexts[cb_type]);
760 }
761 
762 void
763 taskqueue_thread_loop(void *arg)
764 {
765 	struct taskqueue **tqp, *tq;
766 
767 	tqp = arg;
768 	tq = *tqp;
769 	taskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT);
770 	TQ_LOCK(tq);
771 	while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) {
772 		/* XXX ? */
773 		taskqueue_run_locked(tq);
774 		/*
775 		 * Because taskqueue_run() can drop tq_mutex, we need to
776 		 * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the
777 		 * meantime, which means we missed a wakeup.
778 		 */
779 		if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0)
780 			break;
781 		TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0);
782 	}
783 	taskqueue_run_locked(tq);
784 	/*
785 	 * This thread is on its way out, so just drop the lock temporarily
786 	 * in order to call the shutdown callback.  This allows the callback
787 	 * to look at the taskqueue, even just before it dies.
788 	 */
789 	TQ_UNLOCK(tq);
790 	taskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN);
791 	TQ_LOCK(tq);
792 
793 	/* rendezvous with thread that asked us to terminate */
794 	tq->tq_tcount--;
795 	wakeup_one(tq->tq_threads);
796 	TQ_UNLOCK(tq);
797 	kthread_exit();
798 }
799 
800 void
801 taskqueue_thread_enqueue(void *context)
802 {
803 	struct taskqueue **tqp, *tq;
804 
805 	tqp = context;
806 	tq = *tqp;
807 	wakeup_one(tq);
808 }
809 
810 TASKQUEUE_DEFINE(swi, taskqueue_swi_enqueue, NULL,
811 		 swi_add(NULL, "task queue", taskqueue_swi_run, NULL, SWI_TQ,
812 		     INTR_MPSAFE, &taskqueue_ih));
813 
814 TASKQUEUE_DEFINE(swi_giant, taskqueue_swi_giant_enqueue, NULL,
815 		 swi_add(NULL, "Giant taskq", taskqueue_swi_giant_run,
816 		     NULL, SWI_TQ_GIANT, 0, &taskqueue_giant_ih));
817 
818 TASKQUEUE_DEFINE_THREAD(thread);
819 
820 struct taskqueue *
821 taskqueue_create_fast(const char *name, int mflags,
822 		 taskqueue_enqueue_fn enqueue, void *context)
823 {
824 	return _taskqueue_create(name, mflags, enqueue, context,
825 			MTX_SPIN, "fast_taskqueue");
826 }
827 
828 static void	*taskqueue_fast_ih;
829 
830 static void
831 taskqueue_fast_enqueue(void *context)
832 {
833 	swi_sched(taskqueue_fast_ih, 0);
834 }
835 
836 static void
837 taskqueue_fast_run(void *dummy)
838 {
839 	taskqueue_run(taskqueue_fast);
840 }
841 
842 TASKQUEUE_FAST_DEFINE(fast, taskqueue_fast_enqueue, NULL,
843 	swi_add(NULL, "fast taskq", taskqueue_fast_run, NULL,
844 	SWI_TQ_FAST, INTR_MPSAFE, &taskqueue_fast_ih));
845 
846 int
847 taskqueue_member(struct taskqueue *queue, struct thread *td)
848 {
849 	int i, j, ret = 0;
850 
851 	for (i = 0, j = 0; ; i++) {
852 		if (queue->tq_threads[i] == NULL)
853 			continue;
854 		if (queue->tq_threads[i] == td) {
855 			ret = 1;
856 			break;
857 		}
858 		if (++j >= queue->tq_tcount)
859 			break;
860 	}
861 	return (ret);
862 }
863